Jure Leskovec 讨论 Kumo 的关系基础模型，该模型无需训练即可在任何数据库上进行准确预测，以及斯坦福大学的 AI 虚拟细胞项目，构建用于人类细胞和生物医学数据的基础模型。
Jure Leskovec discusses Kumo's relational foundation model that makes accurate predictions on any database without training, and Stanford's AI Virtual Cell project building foundation models for human cells and biomedical data.
要点 · TL;DR
关系深度学习直接在原始多表数据上学习，无需手动特征工程。 Relational deep learning learns directly on raw multi-table data without manual feature engineering.
预训练的关系基础模型无需训练即可在任何数据库上进行预测。 Pre-trained relational foundation models predict on any database without training.
关系基础模型在基准测试上比监督模型高出 5-12%。 Relational foundation models outperform supervised models by 5-12% on benchmarks.
核心观点 · Key points
关系深度学习使神经网络能直接在原始多表数据上学习，无需手动特征工程。 Relational deep learning enables neural networks to learn directly on raw multi-table data without manual feature engineering.
预训练的关系基础模型能在任何数据库和任务上做出准确预测，无需训练。 Pre-trained relational foundation models can make accurate predictions on any database and task without training.
科学AI从原始数据构建代表细胞、患者和分子的基础模型。 AI for science builds foundation models representing cells, patients, and molecules from raw data.
关系模型中的上下文学习使用带标签的子图，通过单次前向传播预测新实体。 In-context learning in relational models uses labeled subgraphs to predict on new entities in a single forward pass.
关系基础模型在基准测试上比监督模型提升5-12%，并带来商业影响。 Relational foundation models outperform supervised models by 5-12% on benchmarks and deliver business impact.
反共识 · Contrarian takes
单表表格问题已解决；真正的挑战是多表关系数据。 Single-table tabular problems are solved; the real challenge is multi-table relational data.
关系模型通过可微注意力和LLM摘要比树模型更具可解释性。 Relational models are more explainable than tree-based models via differentiable attention and LLM summaries.
基础模型因预训练在少量数据上表现良好，无需大数据集。 Foundation models work well with little data due to pre-training, not requiring large datasets.
编码代理因细微错误在复杂数据科学任务上失败；需要更高级的API。 Coding agents fail on complex data science tasks due to subtle mistakes; higher-level APIs are needed.
关系模型可部署在大型区块链图和小型销售团队数据上。 Relational models can be deployed on both massive blockchain graphs and small sales team data.
本期章节 · Chapters（共 22）
关系基础模型突破Relational Foundation Model Breakthrough
引言与研究焦点Introduction and Research Focus
从社交网络到生物网络From Social Networks to Biological Networks
关系深度学习 vs 传统MLRelational Deep Learning vs. Traditional ML
关系深度学习的优势Benefits of Relational Deep Learning
预测示例Examples of Predictions
与传统表格ML对比Comparison with Traditional Tabular ML
单表 vs 多表问题Single-table vs. multi-table problems
用关系基础模型解释上下文学习Explaining in-context learning with relational foundation models
数据库单元格上的注意力机制Attention mechanism over database cells
数据需求与预训练Data requirements and pre-training
系统 vs 模型与上下文示例生成System vs model and in-context example generation
基准测试表现与实际影响Performance on benchmarks and real-world impact
冷启动定义与部署示例Cold start definition and deployment examples
部署与用例Deployment and Use Cases
GPU计算与模型效率GPU compute and model efficiency
局限性与瓶颈Limitations and bottlenecks
后训练与微调价值Post-training and fine-tuning value
注意力与数据的优势Benefits of attention and data
对智能体与编码挑战的兴奋Excitement about agents and coding challenges
任务描述与智能体友好APITask description and agent-friendly APIs